[simon1hofmann]

🤖 AI text below 🤖

Description

Reduce QCO → QC conversion overhead by applying rewrites immediately, releasing replaced operations and SSA uses as conversion progresses. Explicit legality for unchanged classical types avoids unnecessary folding; remove the unused branch conversion pattern and its transform dependency.

Regression coverage checks classical folding and unsupported quantum consumers. Existing target-pipeline tests cover rejection of unstructured control flow.

Benchmarks

Fresh native C++ measurements of 4916b0c1d against the conversion implementation at base 3e8ba8b87. The earlier Mac harness and frozen imports were not available in this workspace; these runs use pinned Benchpress OpenQASM inputs and reconstructed targets on DGX Spark.

Workload / timed scope	Runs per variant	Baseline	PR head	Time change
2,000,000 H gates: QCO → QC	3	3.839 s	2.661 s	-30.7%
BWT n37 / line: full native compilation	1	277.05 s	269.98 s	-2.6%
BWT n37 / line: QCO → QC phase	1	60.30 s	53.24 s	-11.7%
QuGAN n395 / heavy-hex: full native compilation	6	6.926 s	7.211 s	+4.1%
QuGAN n395 / heavy-hex: QCO → QC phase	6	0.0906 s	0.0852 s	-6.0%

Values are medians; BWT has only one pair. All matched output QC hashes and gate counts are identical. QuGAN's full-compilation median increased 4.1% despite its faster conversion phase; full times ranged from 6.28–7.86 s for baseline and 7.04–7.39 s for the PR. These measurements do not establish a general end-to-end speedup.

BWT whole-process peak RSS was 12.18 GiB for baseline and 8.01 GiB for the PR, including untimed import and serialization.

Method: DGX Spark pinned to ten Cortex-X925 cores; Clang 23, -O3, ThinLTO, mold, LLVM/MLIR 23.1.0 with assertions enabled. Both executables use the same harness and common libraries; only the QCO-to-QC implementation differs. Serial fresh processes alternate variant order across pairs. The full timer includes QC copying, QC → QCO, target compilation, and return to QC, including API verification. Import, target setup, output verification, serialization, and hashing are outside the timer.

Inputs and topology recipe: Benchpress b695f30e8. BWT uses a symmetric 37-site line; QuGAN uses a symmetric 409-site heavy-hex graph. Native basis: id, sx, x, rz, cz, measurement, reset. QIR Base payload; seed 42, 10 mapping trials, 1 refinement iteration, lookahead 20. The synthetic case has one dynamic qubit and two million consecutive H gates; its timer includes conversion-boundary verification.

Validation

• Exact head 4916b0c1d: 158 native conversion tests and 271 compiler tests passed.
• Repository lint, whole-file C++ lint, and git diff --check passed.
• Hosted CI passes for this head.

Codex assisted with implementation, review, tests, benchmark reruns, and this description. Human review remains required.

Checklist

• The pull request only contains commits that are focused and relevant to this change.
• I have added appropriate tests that cover the new/changed functionality.
• I have updated the documentation to reflect these changes.
• The changes follow the project's style guidelines and introduce no new warnings.
• The changes are fully tested and pass the CI checks.
• I have reviewed my own code changes.

If PR contains AI-assisted content:

• Any agent that created, edited, or submitted GitHub content was explicitly authorized for that scope, as required by our AI Usage Guidelines.
• Every agent-authored or agent-edited public text body begins with the visible disclosure 🤖 *AI text below* 🤖 (titles are exempt).
• I have disclosed AI assistance in the PR description.
• I confirm that I have personally reviewed and understood all AI-generated content, and accept full responsibility for it.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[burgholzer]

I'll rebase and review this next 😌

[simon1hofmann]

I'll rebase and review this next 😌

Sounds good. Me and @MatthiasReumann had some further ideas for improving the conversions quite substantially, but would create follow-ups for those.

[burgholzer]

Trimmed this down a bit while preserving the main speedup. Let's get this in.